Learning to Classify Cough Sounds with Dictionary-Guided Meta-Learning
Ngan Dao Hoang, Dat Tran-Anh, Tien-Dung Do, Cong Tran, Cuong Pham · 2023
Cough is a popular symptom and cough types might usually be associated with some respiratory diseases. Identifying cough types provides a potential contribution in diagnosis and treatment of respiratory diseases such as Chronic Obstructive Pulmonary Disease (COPD) or pneumonia. However, automatically classifying cough is challenging due to the scarcity of labelled data. In this work, we investigate the task of training a cough classification model with a few available labelled coughing data. Particularly, we propose a meta-learning method that can learn and represent feature maps from the cough sounds to build a dictionary of feature vectors for cough classification. Through our extensive experiment on COVID-19 Thermal Face & Cough dataset, the experimental results can achieve up to 91% averaged F1-Score, which surpasses several models on the same dataset. This demonstrates the feasibility of cough classification even labelled data scarcity.